REC-YoloPose: a lightweight model for enhancing human pose estimation performance in multi-scale and complex scenes

Human pose estimation is a computer vision research area, but it faces challenges in balancing model complexity and accuracy. To address this problem, this study proposes an improved model named REC-YoloPose, based on Yolov8sPose. Firstly, the contextual guidance (CG block) is employed to replace traditional convolution, and efficient local attention (ELA) is introduced into the backbone, enhancing the model's feature extraction capability. Secondly, inspired by Repvit, the original Cross-Stage Partial fusion module (C2f) is improved, striking a balance between model parameters and recognition accuracy. Experimental results demonstrate that the proposed model achieves AP50 scores of 93.1% and 87.0% on Leeds sports pose (LSP) dataset and common objects in context (COCO) dataset respectively. Compared with other mainstream pose estimation algorithms, this model reduces computational parameters by 16.9% to 80.5% while maintaining high detection accuracy. Finally, REC-YoloPose is applied to human posture classification, showcasing its practical value in real-world tasks.